llm-content-creator 0.1.0

package/src/domain/workflow/nodes/CheckImageNode.ts

@@ -0,0 +1,325 @@
+/**
+ * CheckImage Node - 配图质检节点
+ *
+ * 对生成的配图进行质量检查
+ */
+
+import { BaseNode } from './BaseNode.js';
+import type { WorkflowState } from '../State.js';
+import type { QualityReport } from '../State.js';
+import type { QualityCheckDetails } from '../../entities/QualityCheck.js';
+import { enhancedLLMService } from '../../../services/llm/EnhancedLLMService.js';
+import { createLogger } from '../../../infrastructure/logging/logger.js';
+
+const logger = createLogger('CheckImageNode');
+
+/**
+ * LLM 图片质检输出
+ */
+interface ImageQualityCheckOutput {
+  score: number;
+  passed: boolean;
+  details: {
+    relevanceScore: number;
+    aestheticScore: number;
+    promptMatch: number;
+  };
+  fixSuggestions?: string[];
+}
+
+/**
+ * 质检 Prompt 模板
+ */
+const CHECK_IMAGE_PROMPT = `你是一位专业的图片审核专家。请评估以下配图的质量。
+
+【配图信息】
+图片 URL: {imageUrl}
+提示词: {prompt}
+文章主题: {topic}
+
+请从以下维度评估（每项 1-10 分）：
+
+1. **相关性**（relevanceScore）：图片与文章内容/主题的相关性
+2. **美学质量**（aestheticScore）：构图、色彩、清晰度等美学指标
+3. **提示词匹配**（promptMatch）：图片是否符合提示词描述的要求
+
+注意：由于你无法直接看到图片，请基于：
+- 提示词的描述质量
+- URL 中可能包含的信息
+- 与主题的关联性
+
+进行评估。
+
+请以 JSON 格式返回：
+{
+  "score": 8.0,
+  "passed": true,
+  "details": {
+    "relevanceScore": 8.5,
+    "aestheticScore": 7.5,
+    "promptMatch": 8.0
+  },
+  "fixSuggestions": ["建议1"]
+}
+
+评分标准：
+- 9-10 分：优秀，无需改进
+- 7-8 分：良好，可以使用
+- 5-6 分：一般，建议优化
+- 1-4 分：较差，需要重新生成
+
+注意：
+1. 只返回 JSON，不要有其他内容
+2. 如果分数低于 7 分，提供具体的改进建议
+`;
+
+/**
+ * CheckImage Node 配置
+ */
+interface CheckImageNodeConfig {
+  minPassingScore?: number;
+  scoreWeights?: {
+    relevance: number;
+    aesthetic: number;
+    promptMatch: number;
+  };
+}
+
+/**
+ * CheckImage Node 实现
+ */
+export class CheckImageNode extends BaseNode {
+  private config: CheckImageNodeConfig;
+
+  constructor(config: CheckImageNodeConfig = {}) {
+    super({
+      name: 'checkImage',
+      retryCount: 2,
+      timeout: 60000, // 60 秒超时
+    });
+
+    this.config = {
+      minPassingScore: 7.0,
+      scoreWeights: {
+        relevance: 0.4,
+        aesthetic: 0.3,
+        promptMatch: 0.3,
+      },
+      ...config,
+    };
+  }
+
+  /**
+   * 调用 LLM 评估单张图片
+   */
+  private async evaluateImage(
+    imageUrl: string,
+    prompt: string,
+    topic: string
+  ): Promise<ImageQualityCheckOutput> {
+    logger.debug('Evaluating image', {
+      imageUrl: imageUrl.substring(0, 50) + '...',
+      prompt: prompt.substring(0, 50),
+    });
+
+    // 1. 构建 Prompt
+    const checkPrompt = CHECK_IMAGE_PROMPT.replace('{imageUrl}', imageUrl)
+      .replace('{prompt}', prompt)
+      .replace('{topic}', topic);
+
+    // 2. 调用 LLM
+    const systemMessage =
+      '你是一位专业的图片审核专家。请严格按照 JSON 格式返回。';
+
+    const result = await enhancedLLMService.chat({
+      messages: [
+        { role: 'system', content: systemMessage },
+        { role: 'user', content: checkPrompt },
+      ],
+      taskId: '', // 这里没有 taskId，使用空字符串
+      stepName: 'checkImage',
+    });
+
+    // 3. 解析 JSON 响应
+    let output: ImageQualityCheckOutput;
+    try {
+      let content = result.content.trim();
+      if (content.startsWith('```json')) {
+        content = content.slice(7);
+      }
+      if (content.startsWith('```')) {
+        content = content.slice(3);
+      }
+      if (content.endsWith('```')) {
+        content = content.slice(0, -3);
+      }
+      content = content.trim();
+
+      output = JSON.parse(content);
+    } catch (error) {
+      logger.error('Failed to parse LLM output as JSON', {
+        content: result.content.substring(0, 500),
+        error: error instanceof Error ? error.message : String(error),
+      });
+
+      // 解析失败时，返回默认评分（中等）
+      logger.warn('Using default score due to parse failure');
+      return {
+        score: 7.0,
+        passed: true,
+        details: {
+          relevanceScore: 7.0,
+          aestheticScore: 7.0,
+          promptMatch: 7.0,
+        },
+        fixSuggestions: [],
+      };
+    }
+
+    // 4. 验证输出
+    if (typeof output.score !== 'number' || output.score < 1 || output.score > 10) {
+      logger.warn('Invalid score, using default', { score: output.score });
+      output.score = 7.0;
+    }
+
+    return output;
+  }
+
+  /**
+   * 执行质检逻辑
+   */
+  protected async executeLogic(state: WorkflowState): Promise<Partial<WorkflowState>> {
+    logger.info('Starting image quality check', {
+      taskId: state.taskId,
+      imageCount: state.images?.length || 0,
+      retryCount: state.imageRetryCount,
+    });
+
+    try {
+      // 1. 检查是否有图片
+      if (!state.images || state.images.length === 0) {
+        logger.warn('No images to check', {
+          taskId: state.taskId,
+        });
+
+        // 没有图片，返回空质检报告
+        return {
+          imageQualityReport: {
+            score: 0,
+            passed: false,
+            hardConstraintsPassed: false,
+            details: {},
+            checkedAt: Date.now(),
+          },
+        };
+      }
+
+      // 2. 评估所有图片
+      const imageReports = await Promise.all(
+        state.images.map(async (image) => {
+          try {
+            return await this.evaluateImage(
+              image.url,
+              image.prompt,
+              state.topic
+            );
+          } catch (error) {
+            logger.error('Failed to evaluate image', {
+              imageUrl: image.url,
+              error: error instanceof Error ? error.message : String(error),
+            });
+
+            // 评估失败时，返回默认评分
+            return {
+              score: 7.0,
+              passed: true,
+              details: {
+                relevanceScore: 7.0,
+                aestheticScore: 7.0,
+                promptMatch: 7.0,
+              },
+              fixSuggestions: [],
+            };
+          }
+        })
+      );
+
+      // 3. 计算平均分
+      const totalScore = imageReports.reduce((sum, r) => sum + r.score, 0);
+      const avgScore = totalScore / imageReports.length;
+
+      // 4. 收集所有改进建议
+      const allFixSuggestions = imageReports.flatMap(
+        (r) => r.fixSuggestions || []
+      );
+
+      // 5. 构建质检详情
+      const details: QualityCheckDetails = {
+        imageScores: imageReports.map((r, index) => ({
+          imageIndex: index,
+          score: r.score,
+          details: r.details,
+        })),
+      };
+
+      // 6. 判断是否通过
+      const passed = avgScore >= this.config.minPassingScore!;
+
+      // 7. 构建质检报告
+      const qualityReport: QualityReport = {
+        score: avgScore,
+        passed,
+        hardConstraintsPassed: passed, // 图片质检没有硬规则
+        details,
+        fixSuggestions: allFixSuggestions,
+        checkedAt: Date.now(),
+      };
+
+      logger.info('Image quality check completed', {
+        taskId: state.taskId,
+        passed,
+        avgScore,
+        imageCount: imageReports.length,
+        suggestionsCount: allFixSuggestions.length,
+      });
+
+      // 如果质检失败，递增重试计数器
+      const result: Partial<WorkflowState> = {
+        imageQualityReport: qualityReport,
+      };
+
+      if (!passed) {
+        result.imageRetryCount = (state.imageRetryCount || 0) + 1;
+        logger.info('Incremented image retry count', {
+          taskId: state.taskId,
+          previousCount: state.imageRetryCount || 0,
+          newCount: result.imageRetryCount,
+        });
+      }
+
+      return result;
+    } catch (error) {
+      logger.error('Image quality check failed', {
+        taskId: state.taskId,
+        error: error instanceof Error ? error.message : String(error),
+      });
+
+      throw error;
+    }
+  }
+
+  /**
+   * 验证输入状态
+   */
+  protected validateState(state: WorkflowState): void {
+    super.validateState(state);
+
+    // 图片质检可以在没有图片的情况下执行（返回空报告）
+    // 所以这里不做强制检查
+  }
+}
+
+/**
+ * CheckImage Node 单例导出（默认配置）
+ */
+export const checkImageNode = new CheckImageNode();